Claude Code Hackathon Winners 2026 — How a Lawyer, a Doctor, and a Musician Beat 13,000 Developers

Anthropic’s “Built with Opus 4.6” Claude Code hackathon drew 13,000 applicants. 500 were selected. 277 submitted projects. The winners? A personal injury lawyer, an interventional cardiologist, an electronic musician, a road technician from Uganda, and one professional software engineer.

This isn’t a fluke. It’s a signal about where software development is heading.

The Results at a Glance

Place	Project	Creator	Background	Prize
🥇 1st	CrossBeam	Mike Brown	Personal Injury Lawyer	$25,000 credits
🥈 2nd	Elisa	Jon McBee	Software Engineer (25yr)	$10,000 credits
🥉 3rd	postvisit.ai	Dr. Michal Nedoszytko	Interventional Cardiologist	$5,000 credits
🎨 Creative	Conductr	Asep Bagja Priandana	Electronic Musician	$5,000 credits
🏗️ Impact	Road Assessment	Kyeyune Kazibwe	Road Technician (Uganda)	Honorable mention

🥇 CrossBeam: AI-Powered ADU Permit Review

The problem: California has issued 429,000+ Accessory Dwelling Unit permits since 2018. Over 90% come back with corrections — not engineering failures, but bureaucratic errors. Wrong code citations. Missing signatures. Local rules overriding state undocumented rules. The average 6-month permit delay costs homeowners $30,000.

The solution: Mike Brown, a lawyer with zero prior software shipping experience, built an AI agent that:

Reads architectural plans via vision AI
Parses city correction letters item by item
Cross-references against California ADU law (Government Code §§ 66310-66342)
Researches city-specific municipal codes via live web search across 480+ cities
Generates professional response packages with proper code citations

Result: Permit corrections go from weeks to 15 minutes, costing ~$3 in API credits.

CrossBeam Tech Stack

Browser (Next.js)
  ↓ API + Supabase Realtime
Cloud Run (Orchestrator)
  ↓ launches isolated sandboxes
Vercel Sandbox (Agent SDK + Opus 4.6 + Skills)
  ↓ reads/writes
Supabase (Database, Realtime, Storage)

Key architecture decisions:

Cloud Run for the orchestrator — serverless functions timeout at 60-300s, but agent runs take 10-30 minutes
Vercel Sandbox for isolated execution with filesystem access (needed for Agent SDK’s claude_code preset)
13 custom skills with 28 reference files covering CA ADU law
Decision tree router loads only 3-5 relevant files per query (not all 28) — lot type → construction type → modifiers → process
Multi-agent: rolling window of 3 concurrent subagents for plan page analysis
“Targeted viewing” — read corrections first, then look at only relevant plan pages. Dropped processing from 35 min → 15 min

Links: GitHub | Live Demo

🥈 Elisa: Visual Block IDE for Kids → Real Deployed Code

The origin: Jon McBee’s 12-year-old daughter needed to program microcontrollers for a 7th-grade invention fair. Jon has 25 years of engineering experience and recently had his workflow transformed by Claude Code. He wanted his daughter to have the same power — but a terminal isn’t designed for kids.

The solution: A block-based visual IDE where you snap spec blocks together and AI agents build real code:

7 visual primitives: Goals, Requirements, Minions, Skills, Rules, Portals, Deployments
Kids describe what they want by snapping blocks
AI agents plan, code, test, and deploy in real time
Full ESP32 hardware integration (sensor data → web dashboard)
Teaches programming concepts as the agents work

The numbers:

Built solo in one week with Claude Code
Zero lines of code written by hand
40,000+ lines of generated code
1,500+ tests

Key insight: “Very soon, the primary artifacts that software creators carry will no longer be source code — they’ll be well-defined specs and tests. Elisa makes that future accessible to the next generation.”

Links: GitHub | Video Demo

🥉 postvisit.ai: AI Patient Follow-Up for Cardiac Care

The problem: Patients leave cardiology appointments confused about their diagnosis, medications, and next steps. Follow-up is inconsistent and physician visibility between appointments is minimal.

The solution: Dr. Michal Nedoszytko, a practicing interventional cardiologist from Poland, built a platform that:

Explains diagnoses in plain language
Analyzes visit notes and AI-scribe transcripts
Surfaces relevant clinical evidence from scientific resources
Provides full EHR integration with HIPAA compliance
Covers both pre-visit and post-visit clinical workflows

The story: Built while coding in hospital hallways, between patients, and on planes. An MD + PhD who had never shipped production software before — placed 3rd out of 13,000 applicants.

Links: postvisit.ai

🎨 Conductr: AI Band That Plays With You in Real Time

The inspiration: Korg KARMA (Kay Algorithmic Realtime Music Architecture) generates musical patterns in real time — drum patterns, bass lines, melodic phrases — but configuring 400+ parameters per effect is nearly impossible for most musicians.

The solution: Asep Bagja Priandana, an electronic musician, built a browser-based MIDI instrument where you play chords on a controller and Claude Opus 4.6 conducts a generative band around you.

Conductr Architecture

Core engine (libgenseq):

1,134 lines of C, 3.6KB memory, zero runtime heap allocation
4 generators: Drums (Euclidean rhythms), Bass (template-based + scale awareness), Melody (constrained random walk), Harmony (diatonic intervals)
Compiled to WebAssembly for browser deployment
Same C code can compile for ARM Cortex-M (microcontrollers) or wrap into a VST plugin

Three-tier timescale design:

Timescale	Component	Responsibility
15ms	C engine loop	Note generation — rock-solid, deterministic, never waits for AI
4s	Performance analyzer	Extracts musical metrics (pitch range, velocity, note density, rhythm)
8s	Claude Opus 4.6	Returns arrangement decisions (change scale, shift energy, add swing)

Critical design principle: The engine never waits for the AI. Opus operates as a conductor, not a performer — it shapes the next phrase, not the current one. When API is unavailable (slow network, rate limit), a rule-based director takes over. The music never stops.

Key insight: “Separate concerns by timescale. When you have components that run at very different speeds — a 15ms audio loop and an 8-second AI call — don’t make the fast loop wait for the slow one.”

Links: GitHub | Write-up | Video

What This Means

Three patterns emerge from these winners:

Domain expertise is the moat. The lawyer understood permit law. The cardiologist understood patient workflows. The musician understood real-time MIDI. Claude Code amplified knowledge they already had.
Architecture matters more than code volume. CrossBeam’s decision tree router, Conductr’s timescale separation, Elisa’s spec-driven primitives — the wins came from smart system design, not raw output.
The bottleneck is shifting. From “can you write code?” to “can you define the problem clearly?” The winners excelled at structuring problems for agents to solve.

The hackathon proved what many suspected: we’re entering an era where the ability to think clearly about problems matters more than the ability to type syntax. Claude Code didn’t replace these builders — it made their domain expertise executable.